OpenWorker is an open-source, MIT-licensed desktop AI coworker announced by Andrew Ng and Rohit Prasad on 23 July 2026. It runs locally on macOS 12+ (Apple Silicon), with a Windows 10/11 x64 build available but not yet code-signed, and delivers finished work — a polished document, a sent Slack message, an updated calendar entry — rather than chat replies. It is a Tauri and React desktop shell over a local Python agent server, and works with your own API keys.

What is the difference between OpenWorker and OpenClaw Launch?

OpenWorker runs on your own desktop and works with your local files and connected work apps, so it stops when your machine does. Colleagues can reach it by @mentioning it in Slack, but the session still opens on your machine with your keys. OpenClaw Launch hosts an always-on agent in a cloud container that a whole team can reach from Telegram, Discord, WhatsApp, WeChat, Slack or web chat, with persistent memory and scheduled jobs, from $3 the first month and $6/mo after.

Can OpenWorker send messages on Telegram or WeChat?

Slack yes, the others no. You can @mention OpenWorker in a Slack channel and it replies in thread, with the session running on your desktop. Its other connectors cover work tools such as Gmail, Outlook, Google Calendar, GitHub, Jira, Notion, Linear, HubSpot and monday.com. Consumer messaging channels such as Telegram, WhatsApp, WeChat and Feishu are covered by a deployed OpenClaw or Hermes agent instead.

Which models does OpenWorker support?

OpenWorker is model-independent and works with your own keys for OpenAI, Anthropic, Google Gemini, DeepSeek, Kimi, GLM, Qwen, MiniMax, Mistral, Grok and Inkling, open-weight models via Together and Fireworks, and local models through Ollama so data stays on your machine.

Can I use OpenWorker and OpenClaw Launch together?

Yes, and they complement each other. Use OpenWorker on your Mac for deep local file work, where in its default interactive mode each consequential action pauses for your approval, and use an OpenClaw Launch agent as the always-on presence that watches channels overnight, runs scheduled jobs and answers your team at any hour.

← All Comparisons

Comparison

OpenClaw Launch vs OpenWorker

Q: Is OpenWorker free?

Yes. OpenWorker is MIT-licensed and free to download and run. You bring your own model API key and pay that provider directly, or run a local model through Ollama at no cost. The only cloud component OpenWorker itself runs is a small service that brokers OAuth handshakes for connectors, which you can skip by entering API keys manually — though your task content still goes to whichever model provider you choose, and to the connectors you enable, unless you run locally through Ollama with no connectors.

OpenWorker is Andrew Ng's open-source desktop AI coworker, announced 23 July 2026 — it doesn't just chat, it hands you the finished document, sends the Slack message, updates the calendar entry. It runs on your Mac with your own API keys. OpenClaw Launch puts an always-on agent in the channels your team already uses, hosted and running whether or not your laptop is open. Different shapes, and they compose better than they compete.

OpenClaw Launch vs OpenWorker at a Glance

Feature	OpenClaw Launch	OpenWorker
Type	Managed multi-channel AI agent platform	Local-first desktop AI coworker
Where it runs	Cloud container, 24/7	Your own machine — macOS 12+ (Apple Silicon); a Windows 10/11 x64 build is downloadable but not yet code-signed
Runs while your laptop is closed	Yes	No — the agent server runs on your desktop
Where you talk to it	Telegram, Discord, WhatsApp, WeChat, Slack, Feishu, web chat, 12+	Desktop app window, or @mention it in Slack — which opens a session on your desktop
Who can use one instance	A whole team, group, or community	Mainly you — colleagues can @mention it in Slack, but it still runs on your machine, with your keys
Setup	~30 seconds, no install	Download the app, add your API key, connect tools
Model choice	20+ frontier models, 100+ via OpenRouter, or BYOK	BYOK across OpenAI, Anthropic, Gemini, DeepSeek, Kimi, GLM, Qwen, MiniMax, Mistral, Grok, Inkling, or local Ollama
Cost of the software	$3 first month, then $6/mo Lite or $20/mo Pro	Free — you pay your model provider directly
Where your data sits	Your managed container (or self-host the framework)	Your machine, except what goes to your LLM provider and the connectors you enable
Tools & integrations	Thousands of skills, MCP, 1,000+ app connections	25+ connectors (GitHub, Slack, Jira, Notion, Linear, HubSpot, Outlook, Gmail, Google Calendar, monday.com), terminal, local files, MCP
Approvals before risky actions	Configurable per instance	Built in — tool calls are risk-classified and pause for you by default (full-access mode available)
Scheduled work	Cron jobs on an always-on container	Automations, but only while the app is running
Licence	OpenClaw framework open source; platform managed	MIT, fully open source
Maturity	Production, thousands of deployed instances	Open beta (launched July 2026)

*Lite is $3 for the first month, then $6/mo. Pro is $20/mo. Both include AI credits, or bring your own key.

What OpenWorker Is

OpenWorker is an open-source, local-first desktop AI coworker from Andrew Ng and Rohit Prasad, announced on 23 July 2026 and MIT-licensed at github.com/andrewyng/openworker. Its pitch is a deliberate swipe at chat UIs: ask for an outcome, not an answer. You request a customer brief, a triaged Slack alert, an untangled calendar, or a drafted report, and it works across your local files and connected apps until the artefact exists.

Architecturally it is a Tauri 2 and React desktop shell over a local Python agent server built on aisuite. The agent loop, your conversations, your connector tokens and your model keys all live on your machine. The only cloud component OpenWorker itself runs is a small service that brokers OAuth handshakes for connectors — and you can skip even that by pasting API keys manually. Your prompts and task content still travel to whichever model provider you chose, unless you point it at a local Ollama model. Ng's framing was explicit: an open, privacy-preserving, model-independent option for AI coworkers.

What OpenWorker does well:

Deliverables, not transcripts — the unit of output is a finished document, a sent message, an updated calendar entry
Genuine model independence — bring a key for OpenAI, Anthropic, Gemini, DeepSeek, Kimi, GLM, Qwen, MiniMax, Mistral, Grok or Inkling, or keep everything on-device with Ollama
A real approval layer — tool calls are classified by risk before they run (read-only, local writes, command execution, external actions). In the default interactive mode, anything consequential stops for your sign-off rather than asking forgiveness; you can loosen that with per-tool allowlists, or turn it off entirely with full-access mode if you want speed over supervision
Local-first privacy — your files never leave the machine except through the LLM provider and connectors you deliberately turn on
Free and MIT — no seat pricing, no vendor account required; you pay your model provider and nobody else

Where its scope ends:

It lives on one desktop. Close the lid and the coworker stops — automations and watches only run while the app does
There is no shared instance. Colleagues can reach yours through Slack, but anyone who wants their own coworker installs it themselves, with their own keys and their own connectors
Slack is genuinely two-way — a colleague can @mention it and get the answer back as a thread reply — but the session opens on your desktop, so your machine is still the thing doing the work. There is no native presence in Telegram, Discord, WhatsApp, WeChat or Feishu
macOS 12+ on Apple Silicon is the primary target. A Windows 10/11 x64 build is downloadable, but it is not code-signed yet, so SmartScreen will warn you — which is why the website still lists Windows as coming soon
Open beta, days old, and moving fast — expect rough edges

What OpenClaw Launch Is

OpenClaw Launch is a managed deployment platform for the open-source OpenClaw and Hermes agent frameworks. You configure an agent visually, click deploy, and about thirty seconds later it is live — running continuously in its own container, reachable from Telegram, Discord, WhatsApp, WeChat, Slack, Feishu, web chat and more, with persistent memory, cron jobs, thousands of skills and MCP tools.

Why people run agents here instead:

It never sleeps — the agent answers at 3am, on a schedule, or while you are on a plane
One agent, everyone — a whole Telegram group, Discord server or WeChat chat shares a single instance
Model freedom too — 20+ frontier models built in, 100+ through OpenRouter, or bring your own key
Nothing to install — no Python, no Node, no Rust toolchain, no machine to keep awake
Portable — the framework is open source and your config exports whenever you want it

The Honest Difference

These two are not really fighting over the same job. OpenWorker optimises for proximity: the agent sits where your files, your terminal and your logged-in apps already are, so it can produce a polished artefact from material that never has to be uploaded anywhere — beyond the model provider you picked and the connectors you switched on. OpenClaw Launch optimises for availability: the agent sits where your people already talk, awake all the time, shared by everyone, indifferent to whether any particular laptop is open.

If your work is "turn this folder of documents into a brief by Friday", proximity wins. If it is "our community needs an assistant in Telegram that answers at any hour and remembers what it was told last week", availability wins. Choosing on privacy alone is the common mistake — both let you keep model inference under your own API key, and both are open source at the framework level.

Can You Use Both?

Yes, and it is a genuinely good pairing. Use OpenWorker on your Mac for the deep, file-heavy, single-operator work where its default approval prompt before each consequential step is exactly what you want. Use an OpenClaw Launch agent as the always-on front door — the thing that watches a channel overnight, runs the morning digest cron, and answers your team's questions while you sleep.

A practical split: OpenWorker drafts the weekly report from local source files with you approving each external action; the hosted agent then publishes the summary into Telegram or Slack on schedule and fields follow-up questions from everyone who reads it.

Which Should You Choose?

Choose OpenWorker if you are one person on a Mac, your source material is local files and logged-in SaaS apps, you want every consequential action to pause for approval, and you would rather pay nothing but your model provider.

Choose OpenClaw Launch if you need the agent live 24/7 in Telegram, Discord, WhatsApp, WeChat, Slack or web chat, shared by a team or community, with persistent memory and scheduled jobs — $3 the first month, then $6/mo, and no machine of your own to keep running.

Frequently Asked Questions

Is OpenWorker free?

Yes. OpenWorker is MIT-licensed and costs nothing to download or run. You supply your own model API key and pay that provider directly, or run a local model through Ollama and pay nothing at all.

Does OpenWorker run on Windows?

Partly. macOS 12+ on Apple Silicon is the primary target, and a Windows 10/11 x64 build is available to download — but it is not code-signed yet, so Windows SmartScreen will warn you when you run it. That is why the website and Ng's announcement both still describe full Windows support as coming soon.

Is OpenClaw Launch an OpenWorker alternative?

Only partly, and it is worth being precise. If you want a desktop coworker that reaches into your local files, OpenWorker is the better fit and OpenClaw Launch is not trying to replace it. If what you actually wanted was an agent that stays awake and lives in your team's chat apps, that is the gap OpenClaw Launch fills.

Can OpenWorker post to Telegram, WhatsApp or WeChat?

Not those three. Slack is the exception and it works properly — @mention OpenWorker in a channel and the answer comes back as a thread reply, with the session running on your desktop. The rest of the connector set covers Gmail, Outlook, Google Calendar, GitHub, Jira, Notion, Linear, HubSpot and similar work tools. Consumer messaging channels like Telegram, WhatsApp, WeChat and Feishu are where a deployed OpenClaw or Hermes agent works instead.

Which models can each one use?

Both are deliberately model-independent. OpenWorker supports OpenAI, Anthropic, Gemini, DeepSeek, Kimi, GLM, Qwen, MiniMax, Mistral, Grok and Inkling via your own key, plus Together, Fireworks and local Ollama. OpenClaw Launch ships 20+ frontier models, 100+ through OpenRouter, and accepts your own key on any plan.

Bottom Line

OpenWorker is the most credible open-source answer yet to "why does my AI keep handing me text instead of finished work", and its typed approval layer is a better piece of engineering than most desktop agents ship with. It is bounded by the machine it runs on. OpenClaw Launch is bounded by nothing except its container — always awake, in every channel, shared by the whole team. Run OpenWorker for the deep local work, and a hosted agent for everything that has to be there when you are not.

The Chinese-language version of this comparison is at /zh/compare/openworker.